Flint
Released Oct 21, 2026 · knowledge cutoff 2026-06
- Status
- Unknown
- Location
- United Kingdom
- Type
- LLM
- Modality
- Text
- Context window
- 262K
- Max output
- —
- Speed
- —
- Price ($/MTok in / out)
- — / —
- Cost per task
- —
- Open weights
- Yes
Strengths & weaknesses
Strengths
- Built to run locally: the official 4-bit MLX build fits on a well-specced MacBook, so a coding agent can work fully offline
- Tuned for agentic coding loops (investigating a codebase, fixing bugs, implementing features, writing tests) rather than general chat
- Recursive ships its own MLX INT4 and NVFP4/INT4 GGUF builds, so local setups don't wait on community quantisations
Weaknesses
- No hosted API: you run it yourself or not at all
- Rough day-one tooling: the Transformers modelling code wasn't in the base release and runtimes such as mlx-serve were still adding support
- First model from a brand-new lab with a single hands-on review so far, so its behaviour beyond Recursive's own claims is largely uncharacterised
Recursive's first model, released 2026-10-21 on Hugging Face (RecursiveSI/flint): 39.8B total / ~2.5B active MoE (49 layers, 256 routed experts, 6 per token plus a shared expert), text in / text out, which Recursive says was built largely by its automated 'Eureka Machine' research system. Benchmarks are Recursive's own model-card figures (SWE-bench Verified 74.2%, Terminal-Bench 2.1 54.9%); the card's 'SWE-bench Pro v2' 89.7% is not the SWE-bench Pro public split this site tracks, so sweBenchPro is null, and IFEval 95.1% is also reported. Context is 262,144 tokens shared between prompt and output; no separate max output is documented (card examples use max_tokens 32768). Recursive claims 125+ tok/s on an M4 Pro MacBook Pro and 150+ on a 64GB RTX Spark, which is unverified and not an Artificial Analysis measurement. availability is self-host: no hosted API or OpenRouter listing was found. No GPQA Diamond or HLE figure was found in the excerpts read.
Benchmarks2 of 10 reported
SWE-bench Verified
Terminal-Bench 2.1
MMLU-Proretired
GPQA Diamond
SWE-bench Pro
AIME (latest)retired
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
For developers
API model strings
Not researched
Licence
Apache License 2.0 · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor